What do linguists do?
Alec Marantz
Silver Professor of Linguistics and Psychology, New York University
Co-PI Neuroscience of Language Laboratory, NYU Abu Dhabi
February, 2019
What do linguists do?1
Or, rather, what do linguists do that draws such luminaries as the late
Tom Wolfe2
to complain so publicly about our enterprise?3
I’ll confine my discussion here to
what we call “syntax,” which is at the core of so-called “Generative Grammar,” the way of doing
linguistics that has dominated linguistics departments in the US for the past 50 years or so.
Linguists deeply engaged in the enterprise of syntax feel that it’s obvious what they are doing is
not only fascinating, but that it also has led to a set of interesting and important results.
However, outsiders can be skeptical, and my own view is that this skepticism can leads to
hostility–and bad press –for two reasons. As I will briefly illustrate with some examples from
my own research, the actual work of linguists is esoteric and specific. Not everyone gets excited
about the details of ant colonies, for example, and not everyone may be kept up at night by the
connection between predicate possession in English and adversity causative constructions in
Japanese.4
However, through their study of esoteric phenomena, linguists claim to be telling us
1
Thanks to Wayne O’Neil, Jay Keyser, and Noam Chomsky for comments on an earlier version of this paper, and to
Ellie Abrams for editorial assistance.
2
Wolfe, Tom. The kingdom of speech. Random House, 2016.
3
Of course, Wolfe claims that both Darwin and Chomsky are charlatans in his book, putting linguistics in the same
company as evolutionary biology.
4
Wood, Jim, and Alec Marantz. "The interpretation of external arguments." The verbal domain (2017): 255-278.
2
something fundamental about ourselves: about the structure of our minds and brains. Human
minds and brains cut closer to home than ant colonies, at least at first blush, so there may be
more inherent interest in linguist debunking than persecuting myrmecologists.
Modern syntax finds its roots in work by Noam Chomsky in the 1950’s (Chomsky’s work
itself does not emerge out of the blue but represents a particular intellectual departure from the
position of his teacher, Zellig Harris, as well as the work of Nelson Goodman). Many linguists
were drawn to the Chomskyan program through his book, Syntactic Structures. Within this book
one can find places where, if you’re destined to become a syntactician, you’ll have an “aha”
moment, and where, if you’re destined to be a skeptic, you’ll wonder what’s going on. My guess
is that the transparency of the generative program to those experiencing the “aha” moment is
what leads to an inadequate explication of methodology within linguistics: reading Syntactic
Structures,5
it seems obvious what the game is. But obviousness is, of course, accompanied by a
tide of presuppositions and assumptions, so it’s worth returning to Syntactic Structures to
understand what launched syntactic investigations and what escapes contemporary critics of
generative grammar.
One can think of the sentences of a language as sequences of units that re-occur. These can
be sounds, roughly represented in orthography by letters, or words. With this recognition that
sentences can be (partially) described as sequences of recurring units comes the realization that
not all sequences of these units are possible sentences in a language. “Understand launched what
and generative of grammar,” for example, is what linguists call “word salad.” How might we
describe what rules or generalization separate the strings of words that do represent sentences in
a language from those that don’t; or, since it’s human speakers of a language that provide the
5
Chomsky, Noam. Syntactic Structures. Walter de Gruyter, 2002. [Original edition published in 1957.]
3
sentences and the judgments about whether the strings could be sentences, how do we describe
the knowledge of speakers about the sequences of words that could be sentences? One
possibility is that our knowledge is essentially string-based. We know that certain categories of
words follow certain other categories – nouns like “cat” follow articles like “the” in English. A
mathematically precise formulism for string-centric “grammars” that describe languages in these
terms is called a “finite state” grammar. Chomsky asks the question of whether human
languages are among the finite state languages, that is, those “generated” (or recognized) by
finite state grammars.
The point that Chomsky makes next is what has inspired generations of linguists that
followed. He claims we can look at mundane facts about English to show that English, and thus
human languages in general, is not finite state (p. 22).
“Let S1, S2, S3, … be declarative sentences [not questions] in English. Then we
can have such English sentences as:
(11) (i) If S1, then S2.
(ii) Either S3, or S4
(iii) The man who said that S5, is arriving today.
In (11i), we cannot have ‘or’ in place of ‘then’; in (11ii), we cannot have ‘then’ in
place of ‘or’; in (11iii), we cannot have ‘are’ instead of ‘is’. In each of these
cases there is a dependency between words on opposite sides of the comma….
But between the interdependent words, in each case, we can insert a declarative
sentence S1, S3, S5, and this declarative sentence may in fact be one of (11i-
iii)….”
So, according to Chomsky, English can have sets of nested dependencies of the sort, if…then,
either…or, man…is (as opposed to men…are) without a particular limit on the number of such
4
dependencies, and this type of nested dependency cannot be described with finite state
grammars.
If you’re a kid destined to be a syntactician, you might be immediately intrigued by these
facts and fascinated by the phrase structure formalism that can generate these nested
dependencies, where phrase structure languages are the next step up in formal complexity from
finite state languages. Moreover, we see that what look like unbounded long-distance
dependencies from the standpoint of strings end up as local dependencies when language
operates on phrases, rather than words: although if and then can be separated by an arbitrarily
long sequence of other words, the if clause and the then clause are essentially sisters within a
bigger “mother” constituent, as are the either and the or clauses. This also applies to the singular
or plural subject (man or men) and the singular or plural verb phrase (with is or are). Further
research in the generative tradition into the 1970’s discovered that languages generally traffic in
local relations, where the locality becomes apparent once the proper structures are assigned to
sentences and the nature of the relations is understood. Moreover, if you’re the syntactician to be
reading Syntactic Structures, you see that you can get started in this field with a pencil and paper
(or the 21st
century equivalent) and your knowledge of your own native language; the facts you
need to get going are ones you can quickly generate yourself.
Those not destined to become syntacticians might have a different reaction to Chomsky’s
arguments here, perhaps starting with the observation that Chomsky actually provides no English
sentences. “S1” etc. are variables – placeholders for sentences that the reader needs to make up.
Or, as I was asked by an astute philosopher in my early years as a linguist, “where’s the data?”
The worry here is that the mathematical structure of language revealed by nested dependencies in
these made up sentences is the mathematical structure of an abstraction (math of math), rather
5
than of the knowledge of speakers of English or of the output of this knowledge. The tradition of
syntacticians that followed Chomsky has been to provide actual example sentences in their
arguments – the schematic, variable-filled examples show up in mathematical linguistics work,
not in journals devoted, e.g., to syntax. And if you’re worried about the status of “data” that are
generated by the syntactician himself and evaluated against his/her own intuitions, recent work
has demonstrated the incredible reliability and replicability of linguistic data, evaluated by the
standards of experimental psychology (linguistics is not confronting a replicability crisis of the
sort apparent in Social Psychology, for example).6
Once we start using sentences as the data of interest for syntax, rather than schemata of
sentences, certain objections to generative syntax arise from alternative ideas about what should
be done with these sentences. The analytic methods of generative grammar evolve from those of
an earlier structuralist approach to linguistics. On this approach, grounded in anthropology,
linguistics is analytic and explanatory, rather than predictive. The goal is to analyze a fixed
corpus of sentences from a language by reducing the description of the corpus via rule-based
generalizations. For example, recognizing that many English sentences take the form of Name
Subject followed by Tensed Verb, followed by Name Object, we could reduce the set of
sentences like, “Bill saw John,” “John saw Bill,” “John likes Bill,” …. to a list of names, a list of
tensed verbs, and the Subject-Verb-Object “rule.” The rule in some sense explains the presence
of the sentences in the corpus.
6
Sprouse, Jon, Carson T. Schütze, and Diogo Almeida. "A comparison of informal and formal acceptability
judgments using a random sample from Linguistic Inquiry 2001–2010." Lingua 134 (2013): 219-248, Sprouse, Jon,
and Diogo Almeida. "Assessing the reliability of textbook data in syntax: Adger's Core Syntax." Journal of
Linguistics 48.3 (2012): 609-652.
6
Rather than consider the rule-based analysis of sentences as a way of systematically reducing
a finite corpus, Chomsky pointed to the creative aspect of language use: speakers produce and
understand sentences they haven’t heard before. This acknowledgement that any finite corpus of
sentences represents an accidental sampling of the utterances that speaker can recognize or
produce argues instead for a linguistics as part of cognitive psychology, or rather, an enterprise
that would engage in prediction beyond corpora of uttered or written sentences. One wouldn’t
find sentences like, “If, either the guy who said he’ll complete the job is arriving today, or he’s
sending someone else, then we’ll be finished by the end of the week,” in any corpus. However,
English speakers can construct as well as comprehend them, and they exhibit the [a [ b [c c] b] a]
set of nested dependencies that indicate that we have at least phrase structure grammars in our
heads (the dependencies: [if [either [the man --- is] or] then]).
From prediction of novel data in a given language, generative syntacticians can proceed to
predictions about sentences in novel languages. The logic is direct and inescapable: if we can
predict how English speakers project from the sentences they hear to novel sentences on the
basis of the general structure of language, and if any child from any parents will project English
in the same way, then the general structure of language should predict the particulars of all
languages. Because language as described here is unique to humans, the general structure of
language – which linguists call “universal grammar” – must reflect in some manner what it
means to be human, e.g., genetically. Syntacticians thus exploit what they have learned about
universal grammar to predict what we should see in all languages: for example, all languages
should show the consequences of structures that are not finite state. If a language, then, seems to
show a limit on embedding of dependencies of the sort described above, this limit would need to
7
derive from some additional grammatical constraint on structure, rather than from the basic rules
of that specific language.
Testing predictions from linguistic theory often has a flavor distinct from what one would
expect from an idealized scientific method, in which experiments test competing hypotheses
generated from interesting alternative theories (not that any science in practice conforms to this
idealization). For example, Yining Nie from my research group was exploring consequences of
her approach to causative constructions for French. Causatives, which describe an agent making
something happen or bringing something about, strongly depend on transitivity in her approach,
as well as other approaches in the “generative” tradition. The approach predicts that although,
“J’ai fait embrasser Jean” (I made/had someone unspecified hug/kiss John), should be fine, “J’ai
fair pleurer” (I made someone unspecified cry) should not be, since the resulting structure has no
direct object. However, the same analysis would predict that, “J’ai fait pleuvoir” (I made it rain)
should pattern with the unacceptable intransitive sentence about crying rather than the perfectly
grammatical transitive sentence about hugging. This is incorrect; “J’ai fait pleuvoir” is
acceptable, even though pragmatically marked.
Syntacticians call this type of counterexample to a prediction a “puzzle” rather than a
disconfirmation or a falsification precisely because there isn’t an interesting alternative theory
that made the correct prediction here. Syntacticians live by this type of wrong prediction;
puzzles of this sort drive the field. In this case, the puzzle is: what is the right approach to
“weather verbs” like “pleuvoir” in French such that the right general account of causatives makes
them behave like they have a direct object in causative constructions? If you have linguistic
leanings, you’ll be immediately intrigued by this question: do we know things about weather
verbs cross-linguistically that might provide a solution to this puzzle? (Indeed, we do.) But if
8
you’re not intrigued, you should at least recognize that there is a puzzle here that could spur
investigations of further facts from a variety of languages. Puzzles are counterexamples to
predictive linguistic theories that arise in the absence of an alternative theory that predicted them.
An unsympathetic view of linguistic analysis through puzzle-solving has led to a type of
critique of linguistics that suggests that these investigations are at best a type of glorified Sudoku
and at worst a type of mystical Kabbalah – syntacticians are engaged in an endless search for
patterns in language, where the patterns are disconnected from the knowledge that speakers
acquire about their languages and exploit in speaking and comprehension. Even a cursory
examination of the technical literature undermines this critique: the enterprise is predictive,
productive, and competitive. Linguists predict data they don’t have, the body of empirical
generalizations uncovered by the methodology grows year by year, and alternative accounts of
phenomena are in fact pitted against each other, with the losers no longer viable. Progress in
linguistics is transparently displayed in our major journals; nevertheless, some scientists and
engineers that deal with language still question the legitimacy of the generative linguistic
enterprise.
A particular critique of generative grammar comes from computational engineers. If one is
interested in processing millions of sentences from the web in the service of predicting the most
useful answers to queries on web browsers, for example, one might find that a (compact, useful)
finite state description of English covers most of any random sampling of such sentences. For
some tasks, moving to a (more complex) phrase structure description provides no noticeable
gains, and even models of human reading behavior may not be statistically improved by
choosing phrase structure grammars over finite state grammars, at least over random sets of test
9
sentences.7
If you’re following closely, you should be thinking, what about sentences like, “If,
either the guy who said he’ll complete the job is arriving today, or he’s sending someone else,
then we’ll be finished by the end of the week”? Doesn’t English involve nested long-distance
dependencies of a sort that demand a phrase structure description?
The short answer is, yes, and the computational people that argue with generative
grammarians take various approaches to dealing with what Chomsky claims are essential, basic
facts about English and about languages in general.8
But what should be clear to anyone reading
these attacks on linguists is that computationalists are not engaged in the same scientific
enterprise as linguists. The linguistic enterprise is about the knowledge of language that
underlies everything that a speaker does with his/her language, including not only writing those
web pages that serve as data for computational linguistics, but also understanding and making
judgments about sentences that are carefully constructed by linguists as test cases to decide
between competing theories. For example, suppose our judgments on nested long-distance
dependencies stopped at one level of embedding. That might be evidence against phrases
structure accounts of language structure, if the judgments were directly related to our knowledge
of language, as opposed to being the result of processing overload. Restricting our analyses to
narrow subsets of possible data, such as reading time data over random texts or performing a
“correct” analysis by a computational parsing program of a large random selection of English
texts, doesn’t fit with usual scientific practice. One doesn’t know in advance where crucial data
7
Frank, Stefan L., and Rens Bod. "Insensitivity of the human sentence-processing system to hierarchical
structure." Psychological science 22.6 (2011): 829-834. But see, e.g., Roark, Brian, et al. "Deriving lexical and
syntactic expectation-based measures for psycholinguistic modeling via incremental top-down
parsing." Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1Volume
1. Association for Computational Linguistics, 2009.
8
E.g., Frank, Stefan L., Rens Bod, and Morten H. Christiansen. "How hierarchical is language use?." Proceedings of
the Royal Society B: Biological Sciences 279.1747 (2012): 4522-4531.
10
might come from, thus, it’s necessary to develop interesting competing theories in one’s domain
of inquiry and design the appropriate experiments to decide between conflicting predictions of
the theories. In the case of finite state vs. phrase structure descriptions of English, Chomsky
provided the crucial experiment and data in Syntactic Structures – the relative paucity of [a [b [c
c] b] a] nested long-distance dependencies in any corpora of English hardly speaks to the point.
Another critique of generative grammar arises from an anthropological perspective on
language, one that animated the development of linguistics as a field in the US in the first half of
the 20th
century. Suppose the structure of language is not universal across the human species,
nor is it predetermined by our DNA. Suppose instead that the apparent uniformity in structure
across well-studied languages reflects the fact that such languages are used by highly literate
industrial societies, with phrase structures necessary for the types of functions that such societies
employ a language for. Suppose we were to find a people and culture for which the dictates of
their society demand the use of a simpler language – not just finite state but actually finite, with a
countable number of sentences. That there is such a finite language is the claim of Dan Everett,
famously championed by Tom Wolfe.9
In a nutshell, Dan Everett claims that the Pirahã of Brazil speak a language that lacks
“recursion.” The meaning of “recursion” and its applicability to various structures and proposed
rules in linguistics has been a matter of debate, but in terms of the syntactic structures under
discussion in this paper, all would agree that sentences of arbitrary form (S1 in Syntactic
Structures, e.g.) embedded within sentences of arbitrary form would constitute a type of
recursion. Everett denies that Pirahã grammar generates such structures, nor do Pirahã sentences
9
E.g., Everett, Daniel, et al. "Cultural constraints on grammar and cognition in Pirahã: Another look at the design
features of human language." Current Anthropology 46.4 (2005): 621-646.
11
contain noun phrases (of arbitrary structure) embedded within noun phrases, or adjective phrases
within adjective phrases. In fact, given the constraints Everett proposes for Pirahã, the language
is essentially finite: one could list all the sentences of the language.
Although trained as a generative linguist (and although the author of excellent descriptions of
Pirahã written within the Chomskyan generative tradition, in which Pirahã displays a recursive
grammar of the usual sort), Everett writes his current description of the Pirahã as an
anthropologist. He tries to explain the properties he observes in the Pirahã language by referring
to aspects of Pirahã culture. In particular, he attributes to the Pirahã a commitment to the here
and now, constituting a general cultural constraint against talking or thinking about times or
places in the distance, specifically those outside one’s own experience and the experience of the
current set of tribe members. In his writing, Everett tries to connect this general principle to
various additional aspects of the Pirahã language and culture – lack of color terms, lack of
numbers, etc.
Although I am not an anthropologist, I can say that Everett has written mediocre to poor
anthropology. I can say this because it’s the cultural anthropologist’s job to provide the reader
an understanding of the culture being described – to tell a compelling story that allows us to
understand the difference between another culture and our own – and I don’t get any feeling for
the Pirahã or understanding of their culture from reading Everett. For the present discussion, the
largest failure in Everett’s anthropological explanations is his notion that cultural constraints can
explain behavior strictly adhering to the constraints when the possibility of a violation of the
constraints is apparent. I would buy the cultural constraint explanation better if Everett included
stories about how the tribe dealt with transgression and rebellion. A story about how the group
dealt with, say, a rebellious teenager who code-switched between Pirahã and Portuguese (or who
12
yelled out “large happy fish” in Pirahã, with recursive noun phrase structured (see below)) would
go a long way convincing me that Everett has described an alternative way of being human,
rather than an exotic, inscrutable people beyond our understanding.
But as I claimed earlier, even if a poor anthropologist, Everett is a good linguist. As such, he
knew he needed to provide a (semi)formal account of the grammar of Pirahã – how to describe
both the structure of Pirahã as well as the limited nature of its syntax. His analysis of Pirahã
actually involves claiming Pirahã is just like every other language, except that it has a version of
a mechanism that other languages use that, in Pirahã, limits the level of embedding of words
within phrases. Essentially, Everett claims that every noun, verb, or adjective in Pirahã needs to
be “licensed” in its syntactic position by being in construction with (structurally close enough to)
an “evidential” marker – a morpheme that attributes knowledge of the propositional content of a
sentence to the speaker or to someone else (like “reportedly” in English). There’s essentially one
evidential marker per sentence, at the level of the whole sentence, so any noun, verb, or adjective
must be high enough within the constituents of the whole sentence to be visible for licensing.
The meaning of “restrictive” modification of a noun involves a computation that is semantically
recursive (basic modification is “intersective”: a “green tree” refers to a member of the
intersection of the things that are green with the things that are trees). If a language allows
modification of a noun, then the result of such modification, being of the same semantic type as a
noun, may serve as the input to additional modification. In Pirahã, however, if there’s a single
evidential marker on the verb of a sentence (and one verb per sentence), then recursive
modification puts elements in the phrase containing the modified noun deeper and deeper within
this phrase and farther and farther from the evidential. In a structure with a verb+evidential and
a noun phrase, say, an object, a [large [happy [fish] ] ] would include at least one adjective or
13
noun too far from the evidential for licensing, even though the meaning of such recursive
modification can be computed by the Pirahã, Everett claims. The requirement for licensing by
an evidential, then, limits the degree of modification to one – no recursion allowed.
In short, Everett’s own formal account of Pirahã syntax makes Pirahã conform to
“generative” claims about universal grammar. The limits on recursion and on the “creativity” of
Pirahã speakers with respect to the creation of new sentences follow not from a direct constraint
on the grammar that would make Pirahã finite-state in its essence, but rather from a particular
form of licensing constraint that acts to restrict recursive structure building. However, that is not
to say that most linguists would agree with Everett’s account.10
In addition to pushing an
anthropological stance in his analysis of the Pirahã language, Everett has denied the applicability
of standard linguistic methodology for the study of Pirahã. Standard informant work on
languages involves asking for judgments of acceptability, for example, of novel examples
constructed by the linguist. Everett proposes that such methodology would be inappropriate for
the Pirahã. Rather, he seems to suggest that we return to the structuralist methodology of
working with corpora of utterances from speakers, which is a bit like waiting for sentences like,
“If, either the guy who said he’ll complete the job is arriving today, or he’s sending someone
else, then we’ll be finished by the end of the week,” to emerge from the mouths of English
speakers. That is, Everett refuses to test competing theories about the structure of Pirahã using
the standard scientific methodology of linguistics.
The project envisioned in Syntactic Structures, that of characterizing human cognition by
exploring the mathematical properties of human language, continues today, primarily among
10
See, e.g., Nevins, Andrew, David Pesetsky, and Cilene Rodrigues. "Pirahã exceptionality: A
reassessment." Language (2009): 355-404.
14
computational linguists.11
Although Chomsky convincingly demonstrated that natural language
requires more computational “power” than that of context-free phrase structure grammars, there
are apparently interesting differences among levels of complexity in the range between context
free grammars and the power of Turing Machines, and the exact place of our linguistic
competence in the hierarchy is still unknown or unproven. However, this question does not drive
my own work in Universal Grammar, which is more concerned with the ontology of elements
that are combined to make words and phrases, and in the specifics of the interactions among
these elements, as opposed to the general mathematical characteristics of these interactions. That
makes me somewhat of an “empiricist,” as Chomsky informed me when I was his student,
whether in a neutral or evaluative tone I’m still not sure to this day.
At the moment (Spring semester of 2019), I’m investigating the nature of “locality”
constraints on the interactions of the combinatory elements of language: the “morphemes.” This
investigation is directly connected to Chomsky’s observation from Syntactic Structures I
explained above, that the apparent long-distance relation between “if” and “then” in a conditional
sentence is actually a very local relation between the “if” clause and the “then” clause, once the
hierarchical structure of language is acknowledged. Certain types of interactions between
morphemes require a very local relationship, similar to the connection between “if” and “then.”
For example, some suffixes restrict the meanings of the stems to which they attach. Although
“novel” can refer both to the content of a book (“this novel is very interesting”) or to the book as
an object (“this novel weighs half a pound”), the verbalizing suffix -ize chooses the “content”
reading. One can “novelize” someone’s life, but one can’t “novelize” these sheets of paper into a
11
See, e.g., Graf, Thomas. "Grammar Size and Quantitative Restrictions on Movement." Proceedings of the Society
for Computation in Linguistics (SCiL) 2018 (2018): 23-33.
15
physical book. If we add another suffix to nominalize the verb, to create “novelization,” the final
suffix isn’t local enough to the stem, “novel,” to choose the physical object reading, which is
thus no longer available to further suffixation after the addition of -ize. We can’t talk about the
“novelization” of reams of paper.
These sorts of observations yield interesting insight into a variety of puzzling phenomena.
For example, the prefix re- seems to mean the same thing as “again”: “I re-opened the door,”
seems equivalent to, “I opened the door again.” However, re- only expresses one of the possible
meanings of “again,” that of restoring the previous state of an object, not that of repeating an
action. So, “I re-opened the door,” means, “I opened the door, and it was open before.” As a
result, re- does not attach to verbs expressing activities, rather than changes of state. Although I
can say, “I danced again last night,” I can’t say “I re-danced last night.”
Now the careful reader should be thinking, isn’t “dance” in “re-dance” close enough to re- to
allow it to choose a special “do again” meaning for re-? Re- is right next to dance, after all.
That is, couldn’t I make “re-dance” mean “dance again” if I wanted to, and get other speakers to
agree? The answer is actually, no, you can’t. The hierarchical arrangement of morphemes in
language is called the “syntax” of the language. The same hierarchical syntactic structure is
interpreted (as Chomsky says, “externalized”) in sound/writing/sign on the one hand and in
meaning on the other. So, we can find clues as to the arrangement of morphemes relevant for
meaning in the way that morphemes are pronounced – both the meaning and the pronunciation
are computed from the same structure in highly constrained ways. The “restitutive” (indicating
restored state) re- prefix in English bears stress (compare unstressed re- in “I returned the bottle”
with stressed re- in “I re-turned the knob to the open position”). In English, the stress on a prefix
indicates that it belongs to a separate syntactic domain from the stem to which it attaches; a full
16
analysis of re- would show it belongs to the syntactic domain of the direct object independent of
the verb. Belonging to separate syntactic domains, as indicated by their pronunciation, re- and
dance cannot influence each other semantically as they could if they belonged to the same
domain. Even though they look like they’re next to each other, like “if” and “then,” re- and
dance are in fact not syntactically local in the relevant sense for semantic interpretation, if dance
were to want to choose the repetitive rather than the restitutive reading of re-.
Investigations into the properties of English words hold strong implications for the
architecture of grammar and the nature of language. However, the sorts of locality constraints at
issue here on the interactions of morphemes do not immediately hold obvious implications for
the nature of human cognition, unlike Chomsky’s conclusions that human languages involve
more computational power than context free grammars. My type of research is more standard
cognitive science: investigating the representation and use of language without explicitly making
claims that the nature of linguistic representations and computations are either unique to humans
or unique to language. Or, I should say, my type of research is standard cognitive neuroscience,
since these investigations include testing hypotheses via connections to neural mechanisms. For
example, one implication of the theory of morphemes and their combination that I’ve been
developing on the basis of the sort of evidence about novelization and re-dance that I sketched
above is that English speakers must analyze words like “excursion” into stem and suffix
morphemes “excurse” and “ion” even though the putative stem “excurse” occurs only in the
word “excursion.” My lab has confirmed this, as well as other surprising predictions about the
neural processing of morphologically complex words, using a brain monitoring technique called
“magnetoencephalography,” which measures the magnetic fields generating by electrical activity
17
in neurons in the cerebral cortex.12
Nothing about the linguistic enterprise demands that one
marshal evidence from brain signals to support one hypothesis over another. But, more
crucially, nothing about linguistics tells us that we won’t find crucial evidence for a theory in
brain responses either.
What do linguists do? We’re not engineers and we’re not anthropologists; critiques of
linguistics based on the goals of these disciplines are therefore at best baffling. We are cognitive
neuroscientists and, perhaps with vision scientists, the pioneers of this discipline. We operate at
many levels of abstraction, from the quite general claims about human cognition based on
computational linguistic analysis of the mathematics of human grammars to specific hypotheses
about the neural responses to “excursion.” Like the findings of any science, the observations of
linguists from the behavior of “if…then” sentences to the ungrammaticality of “re-dance” should
appeal to our innate curiosity as humans and to our desire to understand more about the world
around us. Chomsky suggests, following the rationalist tradition of Descartes among others, that
linguistics may hold special appeal to the extent that language represents a species-specific,
genetically encoded capacity; understanding the nature of language is part and parcel of an
appreciation of what makes us human. However, I’d like to believe that Chomsky’s discussion
of nested dependencies in English syntax, or my more humble description of the (im)possible
meanings of novelization and re-dance, would be sufficient to evoke the foundations of a
fascinating and productive field of study.
12
Gwilliams, L., and A. Marantz. "Morphological representations are extrapolated from morpho-syntactic
rules." Neuropsychologia 114 (2018): 77-87.